One solution seems to be to pre-flatten the data in a CTE, resulting in
dramatically lower runtimes:
~~~
WITH flat AS (SELECT id, FLATTEN(data) AS data)
SELECT id, data[0] AS dttm, data)[1] AS result
FROM flat
~~~
This was tested on a single node, and each JSON array to be flattened
has 1,440 elements. I can extract some query profiles, and will enter a
Jira if the plan looks like it could be optimized.
On 19 Jul 2016, at 16:40, rahul challapalli wrote:
Matt,
Having multiple flatten's in your query leads to cross-join between
the
output of each flatten. So a performance hit is expected with the
addition
of each flatten. And there could also be a genuine performance bug for
this
scenario. To be sure it is a bug we need more information as Abhishek
pointed out.
However if you want to do some computations after you flattened out
your
query, it might be helpful sometimes to rewrite the query such that
multiple flatten's fall in multiple sub-queries. You may see some
performance improvement. Let me know how it goes.
- Rahul
On Tue, Jul 19, 2016 at 1:22 PM, Abhishek Girish
<abhishek.gir...@gmail.com>
wrote:
Hi Matt,
Can you please share more information on your setup, specifically the
size
of your dataset, including an approximate average size of individual
JSON
files, the number of nodes, including Drillbit memory config.
Also can you share the query profiles for the few scenarios you
mention.
Regards,
Abhishek
On Friday, July 15, 2016, Matt <bsg...@gmail.com> wrote:
I have JSON data with with a nested list and am using FLATTEN to
extract
two of three list elements as:
~~~
SELECT id, FLATTEN(data)[0] AS dttm, FLATTEN(data)[1] AS result FROM
...
~~~
This works, but each FLATTEN seems to slow the query down
dramatically,
3x
slower with the second flatten.
Is there a better approach to extracting list elements?
~~~
[
{
"id": 16,
"data": [
[
"2016-07-13 00:00",
509,
"OK"
],
[
"2016-07-13 00:01",
461,
"OK"
],
[
"2016-07-13 00:02",
508,
"OK"
],
~~~
